Published on July 13, 2026 at 06:00 CEST (UTC+2)
Since Chromium 148, Math.tanh is now fingerprintable to link underlying OS (376 points by joahnn_s)
Since Chromium 148, Math.tanh is now fingerprintable to link underlying OS – This article reveals that the JavaScript Math.tanh function returns slightly different floating-point results depending on the operating system’s math library (glibc, libsystem_m, UCRT). These differences, often just one unit in the last place (ULP), allow anti-bot systems to fingerprint the underlying OS with high accuracy. The signal is quieter than traditional canvas or WebGL fingerprinting but persists across Chrome versions. This technique can be used to detect automation or distinguish genuine users from bots.
GhostLock, a stack-UAF that has existed in all Linux distributions for 15 years (130 points by ranger_danger)
GhostLock, a stack-UAF that has existed in all Linux distributions for 15 years – The article details CVE-2026-43499, a Linux kernel vulnerability present in all major distributions since 2011. An unprivileged attacker can trigger a use-after-free on kernel stack memory using regular threading syscalls. The exploit was turned into a 97% stable privilege escalation and container escape, earning a $92,337 bug bounty. The write-up covers root cause, exploitation techniques (including ASLR bypass and PR_SET_MM_MAP), and the proposed kernel patches (RANDOMIZE_KSTACK_OFFSET, STATIC_USERMODE_HELPER).
Cyberpunk Comics, Manga and Graphic Novels (109 points by zdw)
Cyberpunk Comics, Manga and Graphic Novels – This is a curated list of illustrated literature spanning Western comics, Japanese manga, and graphic novels, sorted by publication date. It includes classics like The Long Tomorrow (1975) by Dan O’Bannon and Moebius, which heavily influenced later cyberpunk works such as Blade Runner. The article serves as a media index for fans of the cyberpunk genre, providing themes, settings, and publication details for each entry.
Tiny Emulators (175 points by naves)
Tiny Emulators – A web-based collection of emulators for vintage 8-bit computers and consoles, including the ZX Spectrum, Commodore 64, Amstrad CPC, and various Soviet-era machines. The page provides interactive UIs that allow users to run classic software, demos, and games directly in the browser. It is a nostalgic and educational resource for retro computing enthusiasts.
Ask HN: Add flag for AI-generated articles (179 points by levkk)
Ask HN: Add flag for AI-generated articles – A Hacker News user proposes adding a flag to mark articles as AI-generated, not to de-rank them but to let readers skip them. The discussion reveals that HN already bans AI-generated text in comments and users generally discount such articles. Dang (HN moderator) notes that the community’s allergic response is a form of arms race between AI training and human detection, with readers increasingly demanding to see the “prompt” behind AI-written content.
So you want to learn physics (second edition, 2021) (122 points by azhenley)
So you want to learn physics (second edition, 2021) – This is a comprehensive self-study guide for learning physics from scratch, written by Susan Rigetti. It provides a detailed curriculum covering undergraduate and graduate-level physics, with recommended textbooks and resources. The guide has been used by over 600,000 readers, many of whom have gone on to earn physics degrees, while others pursue it out of pure curiosity.
Designing and assembling my first PCB (67 points by tadasv)
Designing and assembling my first PCB – The author recounts their journey from playing with an Arduino Nano ESP32 to designing and assembling their first custom printed circuit board. They describe using LLMs for help with firmware, moving from breadboard prototypes to a “production” workflow. The article highlights the accessibility of modern hardware development tools and the satisfaction of moving beyond pre-built modules.
Why Vanilla JavaScript (94 points by guseyn)
Why Vanilla JavaScript – (Content preview not available, but the title suggests an advocacy for using plain JavaScript without frameworks or libraries, likely arguing for simplicity, performance, and maintainability.)
1970 Plymouth Hemi 'CUDA (4 points by frobinson47)
1970 Plymouth Hemi ‘CUDA – A passionate homage to the iconic muscle car, describing it as a cultural statement that embodies raw power and aggressive design. The article places the Hemi ‘Cuda in the context of the muscle car era, arguing that it represents the pinnacle of Chrysler engineering and a rejection of subtlety.
Ask HN: What Are You Working On? (July 2026) (92 points by david927)
Ask HN: What Are You Working On? (July 2026) – A monthly megathread where Hacker News users share their side projects and ideas. Examples include building custom software for an NGO using AI agents to test domain models, and creating machine-readable databases of laws for Indonesia, Singapore, and Malaysia tailored for AI agents.
Fingerprinting moves beyond traditional vectors into subtle mathematical signals – The Math.tanh vulnerability shows that even low-level floating-point computation can become a fingerprinting vector. As anti-bot systems evolve, any deterministic difference between platforms will be exploited. For ML developers, this means that model inference outputs or preprocessing steps that rely on platform-specific libraries could leak information, raising privacy and security concerns in deployed AI systems.
AI is both a target and a tool in cybersecurity exploits – The GhostLock vulnerability exploit earned a significant bounty, and its write-up discusses techniques like ASLR spray and stack forging. Meanwhile, AI is increasingly used to automate vulnerability discovery and exploit generation. The arms race between offensive AI and defensive AI is accelerating, and kernel security patches must now consider adversarial ML attacks that could bypass randomization techniques.
Community backlash against AI-generated content is reshaping platform policies – The HN discussion about adding an AI-generation flag reflects growing user distrust of AI-written articles. Platforms are experimenting with labels, detection heuristics, and even outright bans. For ML practitioners, this signals a need for transparency: any system that generates text for public consumption should include clear provenance metadata, or risk being ignored or penalized by human communities.
AI agents are being integrated into legal and regulatory domains – The “What Are You Working On” thread includes projects that structure legal texts specifically for AI agents. This trend points to the emergence of “agent-ready” data: machine-readable, semantically tagged datasets that enable LLMs to retrieve and reason over laws reliably. As AI agents become more autonomous, the demand for structured, verifiable domain knowledge will grow, creating opportunities for data curation and ontology design.
LLMs lower the barrier to entry for hardware and firmware development – The PCB design article explicitly credits LLMs for helping with Arduino firmware. This democratization of embedded systems means that AI-assisted coding is moving beyond software into hardware-adjacent fields. Future ML trends will include more training data for low-level programming languages and real-time operating systems, as well as specialized models for circuit design and PCB layout.
Nostalgia and retro computing remain a strong niche for web-based AI performance – The Tiny Emulators page demonstrates that in-browser emulation of 8-bit systems is now feasible and performant. This is partly due to advances in WebAssembly and JavaScript engines, which themselves benefit from compiler optimizations driven by ML (e.g., in V8). AI/ML researchers can use such emulators as cheap, reproducible environments for testing AI agents on classic game tasks or evaluating historical software compatibility.
The “long tail” of user-generated content on HN reveals real-world ML applications – The monthly “What Are You Working On” thread showcases dozens of practical AI projects: NGO operations software with AI-driven scenario testing, legal databases for AI agents, and more. This indicates that the most impactful ML deployments are often bespoke, data-specific, and built by individuals or small teams. For the AI industry, this underscores the importance of low-code AI platforms and accessible model APIs that empower non-experts to build custom solutions.
Analysis generated by deepseek-reasoner